policy            success  calls/question  error calls/question  final context tokens
scripted eps=0.0    1.000            4.06                  0.00                  77.0
scripted+noise eps=0.15    0.925            4.42                  0.56                  78.8
scripted+noise eps=0.4    0.840            5.46                  1.49                  88.7

BM25 recall@3 of the needed fact over 1600 lookups: 0.9956

mean tokens added per call (action + observation): 17.5
 calls T  tokens read, full re-read   new tokens (cached prefix)   ratio
       5                        263                           88     3.0
      10                        964                          175     5.5
      20                       3682                          351    10.5
      40                      14379                          701    20.5
